Study looks at the effect that failed links have on the throughput of HPC systems: What workloads are most effected? How many links need to be down before throughput of the machine is noticeably affected?
The U.S. Army Research Office (ARO), in partnership with IARPA, are investigating innovative, efficient, and scalable computer architectures that are capable of executing next-generation large scale data-analytic applications. These applications are increasingly sparse, unstructured, non-local, and heterogeneous. Under the Advanced Graphic Intelligence Logical computing Environment (AGILE) program, Performer teams will be asked to design computer architectures to meet the future needs of the DoD and the Intelligence Community (IC). This design effort will require flexible, scalable, and detailed simulation to assess the performance, efficiency, and validity of their designs. To support AGILE, Sandia National Labs will be providing the AGILE-enhanced Structural Simulation Toolkit (A-SST). This toolkit is a computer architecture simulation framework designed to support fast, parallel, and multi-scale simulation of novel architectures. This document describes the A-SST framework, some of its library of simulation models, and how it may be used by AGILE Performers.
Abstract not provided.
Proceedings of INDIS 2020: Innovating the Network for Data-Intensive Science, Held in conjunction with SC 2020: The International Conference for High Performance Computing, Networking, Storage and Analysis
Priority-based Flow Control (PFC), RDMA over Converged Ethernet (RoCE) and Enhanced Transmission Selection (ETS) are three enhancements to Ethernet networks which allow increased performance and may make Ethernet attractive for systems supporting a diverse scientific workload. We constructed a 96-node testbed cluster with a 100 Gb/s Ethernet network configured as a tapered fat tree. Tests representing important network operating conditions were completed and we provide an analysis of these performance results. RoCE running over a PFC-enabled network was found to significantly increase performance for both bandwidth-sensitive and latency-sensitive applications when compared to TCP. Additionally, a case study of interfering applications showed that ETS can prevent starvation of network traffic for latency-sensitive applications running on congested networks. We did not encounter any notable performance limitations for our Ethernet testbed, but we found that practical disadvantages still tip the balance towards traditional HPC networks unless a system design is driven by additional external requirements.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Proceedings - IEEE International Conference on Cluster Computing, ICCC
Avoiding communication bottlenecks remains a critical challenge in high-performance computing (HPC) as systems grow to exascale. Numerous design possibilities exist for avoiding network congestion including topology, adaptive routing, congestion control, and quality-of-service (QoS). While network design often focuses on topological features like diameter, bisection bandwidth, and routing, efficient QoS implementations will be critical for next-generation interconnects. HPC workloads are dominated by tightly-coupled mathematics, making delays in a single message manifest as delays across an entire parallel job. QoS can spread traffic onto different virtual lanes (VLs), lowering the impact of network hotspots by providing priorities or bandwidth guarantees that prevent starvation of critical traffic. Two leading topology candidates, Dragonfly and Fat Tree, are often discussed in terms of routing properties and cost, but the topology can have a major impact on QoS. While Dragonfly has attractive routing flexibility and cost relative to Fat Tree, the extra routing complexity requires several VLs to avoid deadlock. Here we discuss the special challenges of Dragonfly, proposing configurations that use different routing algorithms for different service levels (SLs) to limit VL requirements. We provide simulated results showing how each QoS strategy performs on different classes of application and different workload mixes. Despite Dragonfly's desirable characteristics for adaptive routing, Fat Tree is shown to be an attractive option when QoS is considered.
Remote Direct Memory Access (RDMA) over Converged Ethernet (RoCE) has the potential to provide performance that rivals traditional high performance fabrics. If this potential proves out, significant impacts on system procurement decisions could follow. This work provides a series of small scale performance results which are used to compare and contrast the performance of RoCE-enabled Ethernet with TCP-based Ethernet and an HPC network. Additionally, a discussion of the maturity of RoCE firmware/software stacks and documentation is provided along with useful approaches for probing performance. A detailed description of two experimental setups known to have good RoCE performance is given, including step-by-step configuration and the exact hardware and software revisions employed. At small scales, RoCE is found to have significant performance advantages over "out-of-the-box" TCP protocols and is competitive with state-of-the-art high performance networks. Further examination of RoCE using a wider array of benchmarks and at greater scale is warranted.
Network security researchers often rely on EmulyticsTM to provide a way to evaluate the safety and security of real world systems. This work involves running a large number of virtual machines on a distributed platform to observe how software and hardware will respond to different types of attacks. While EmulyticsTM software such as minimega [2] provide a scalable system for conducting experiments, the sheer volume of network traffic produced in an experiment can easily exceed the rate at which data can be recorded for offline analysis. As such, researchers must perform live analytics, narrow their monitoring scope or accept that they must run an experiment multiple times to capture all the information they require. In support of Sandia's commitment to EmulyticsTM, we are developing new storage components for the Carlin cluster that will enable researchers to capture significantly more network traffic from their experiments. This report provides a summary of Haoda Wang's initial investigation of how new AMD Epyc storage nodes can be adapted to perform packet capture at 100Gbps speeds with minimal loss. This work found that the NVMe storage capabilities of the Epyc architecture are suitable for capturing 100Gbps Ethernet traffic. While capturing traffic with existing libraries was surprisingly challenging, we were able to develop a DPDK-based software tool that recorded network traffic to disk with minimal packet loss.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Data movement is considered the main performance concern for exascale, including both on-node memory and off-node network communication. Indeed, many application traces show significant time spent in MPI calls, potentially indicating that faster networks must be provisioned for scalability. However, equating MPI times with network communication delays ignores synchronization delays and software overheads independent of network hardware. Using point-to-point protocol details, we explore the decomposition of MPI time into communication, synchronization and software stack components using architecture simulation. Detailed validation using Bayesian inference is used to identify the sensitivity of performance to specific latency/bandwidth parameters for different network protocols and to quantify associated uncertainties. The inference combined with trace replay shows that synchronization and MPI software stack overhead are at least as important as the network itself in determining time spent in communication routines.
Abstract not provided.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Performance modeling of networks through simulation requires application endpoint models that inject traffic into the simulation models. Endpoint models today for system-scale studies consist mainly of post-mortem trace replay, but these off-line simulations may lack flexibility and scalability. On-line simulations running so-called skeleton applications run reduced versions of an application that generate traffic that is the same or similar to the full application. These skeleton apps have advantages for flexibility and scalability, but they often must be custom written for the simulator itself. Auto-skeletonization of existing application source code via compiler tools would provide endpoint models with minimal development effort. These source-to-source transformations have been only narrowly explored. We introduce a pragma language and corresponding Clang-driven source-to-source compiler that performs auto-skeletonization based on provided pragma annotations. We describe the compiler toolchain, validate the generated skeletons, and show scalability of the generated simulation models beyond 100Â K endpoints for example MPI applications. Overall, we assert that our proposed auto-skeletonization approach and the flexible skeletons it produces can be an important tool in realizing balanced exascale interconnect designs.
Proceedings - 2017 17th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing, CCGRID 2017
The power and procurement cost of bandwidth in system-wide networks has forced a steady drop in the byte/flop ratio. This trend of computation becoming faster relative to the network is expected to hold. In this paper, we explore how cost-oriented task placement enables reducing the cost of system-wide networks by enabling high performance even on tapered topologies where more bandwidth is provisioned at lower levels. We describe APHiD, an efficient hierarchical placement algorithm that uses new techniques to improve the quality of heuristic solutions and reduces the demand on high-level, expensive bandwidth in hierarchical topologies. We apply APHiD to a tapered fat-Tree, demonstrating that APHiD maintains application scalability even for severely tapered network configurations. Using simulation, we show that for tapered networks APHiD improves performance by more than 50% over random placement and even 15% in some cases over costlier, state-of-The-Art placement algorithms.
Abstract not provided.
Abstract not provided.
Proceedings of COM-HPC 2016: 1st Workshop on Optimization of Communication in HPC Runtime Systems - Held in conjunction with SC 2016: The International Conference for High Performance Computing, Networking, Storage and Analysis
We introduce a topology-aware performance optimization and modeling workflow for AMR simulation that includes two new modeling tools, ProgrAMR and Mota Mapper, which interface with the BoxLib AMR framework and the SSTmacro network simulator. ProgrAMR allows us to generate and model the execution of task dependency graphs from high-level specifications of AMR-based applications, which we demonstrate by analyzing two example AMR-based multigrid solvers with varying degrees of asynchrony. Mota Mapper generates multiobjective, network topology-aware box mappings, which we apply to optimize the data layout for the example multigrid solvers. While the sensitivity of these solvers to layout and execution strategy appears to be modest for balanced scenarios, the impact of better mapping algorithms can be significant when performance is highly constrained by network hop latency. Furthermore, we show that network latency in the multigrid bottom solve is the main contributing factor preventing good scaling on exascale-class machines.
Discrete event simulation provides a powerful mechanism for designing and testing new extreme- scale programming models for high-performance computing. Rather than debug, run, and wait for results on an actual system, design can first iterate through a simulator. This is particularly useful when test beds cannot be used, i.e. to explore hardware or scales that do not yet exist or are inaccessible. Here we detail the macroscale components of the structural simulation toolkit (SST). Instead of depending on trace replay or state machines, the simulator is architected to execute real code on real software stacks. Our particular user-space threading framework allows massive scales to be simulated even on small clusters. The link between the discrete event core and the threading framework allows interesting performance metrics like call graphs to be collected from a simulated run. Performance analysis via simulation can thus become an important phase in extreme-scale programming model and runtime system design via the SST macroscale components.
Abstract not provided.
Abstract not provided.